CENTRALIZED LEARNING FOR THE DEEP Q-LEARNING MODELS
Viacheslav BOCHOK, Nataliia FEDOROVA · INFORMATION TECHNOLOGY AND SOCIETY · 2024
The article is devoted to centralized learning and knowledge sharing between Deep Q-leaning agents. Multi- agent systems are fault-tolerant and capable of self-organization, but achieving this can require a lot of resources. The agent independently explores the environment, gradually adapting to different situations. For systems where the state space is continuous, and therefore has many options, and the outcome of the transition in the future is unknown, it is difficult for the agent to choose to explore the space of actions and states, select a more profitable strategy and not get stuck in pseudo-winning strategies (local minima). The goal is to increase the stability of the learning process. On the example of the MADDPG approach and the KnowSR framework, the following methodology was proposed: to use several agents that exchange experience and knowledge between models, forming a common buffer. The scientific novelty is the use of centralized learning to increase the stability of actions of Deep Q learning agents with a mechanism for sharing already learned knowledge.